feat(zenifra): add hosted model catalog

[ramonpaolo]

Summary

• Add the 11 model routes currently hosted by Zenifra.
• Replace the obsolete Zenifra Qwen3.6 route with Qwen3.8 27B and align its capabilities.
• Convert Zenifra BRL prices to models.dev USD/MTok using the project rate R$5.20 = US$1.00.
• Document the conversion rate and date, including the qwen3.7-plus context tier.

Validation

• bun validate passes.
• git diff --check passes.
• bun test: 318 passed, 6 pre-existing failures unrelated to this PR; no test files are changed.

[github-actions]

Action items

• [high] [violation] providers/zenifra/models/deepseek-v4-pro.toml:5 - Check: Relay reasoning_options must follow the lab/same-surface peer baseline for the underlying model, not an invented effort enum. Why: First-party DeepSeek V4 Pro is toggle + high/max (providers/deepseek/models/deepseek-v4-pro.toml); solid peers match that. This file authors only low|medium|high|xhigh|max, which invents graded levels the lab does not expose and drops the on/off control. Action: Set reasoning_options to lab/peer shape (toggle + high/max, with a leading wire comment) unless Zenifra docs or a live request show a different control surface; do not keep a full L/M/H-style enum by default.
• [high] [violation] providers/zenifra/models/deepseek-v4-flash-0731.toml:5 - Check: Same DeepSeek V4 relay baseline. Why: Lab Flash paths use toggle + low/high/max (or high/max); this entry uses low|medium|high|xhigh|max with no toggle, which is not the native/peer set. Action: Align to DeepSeek lab/peer controls (include toggle if the host forwards thinking.type / equivalent); only keep extra levels with host-specific evidence.
• [high] [violation] providers/zenifra/models/glm-5.2.toml:5 - Check: GLM-5.2 effort baseline is lab high/max, not a GPT-style ladder. Why: providers/zhipuai/models/glm-5.2.toml and providers/zai/models/glm-5.2.toml document effective levels high/max (with low/medium mapped to high). This PR sets low|medium|high|xhigh|max, inventing a broader enum. Action: Use [{ type = "effort", values = ["high", "max"] }] (or lab-equivalent with verified none off) unless Zenifra proves different wire values.
• [high] [violation] providers/zenifra/models/qwen3.7-max.toml:4 - Check: Relays must not use reasoning_options = [] when the lab/peers expose caller controls. Why: Alibaba Qwen3.7 Max is toggle + budget_tokens; peers such as Modelis match that. Empty means no control, not “unknown from /v1/models.” Action: Author the host’s real controls (at least lab-like toggle/budget if forwarded), or document affirmative evidence that Zenifra exposes none.
• [high] [violation] providers/zenifra/models/qwen3.7-plus.toml:5 - Check: Same Qwen3.7 Plus relay rule. Why: Lab and peers use toggle + budget_tokens; this file sets []. Action: Copy the lab/peer option set for this host’s API, or prove Zenifra has no thinking controls.
• [high] [violation] providers/zenifra/models/qwen3.8-max.toml:7 - Check: Effort values must be the native/peer set, not a full schema dump. Why: Alibaba Qwen3.8 Max is toggle + low/medium/xhigh + budget_tokens. This PR uses none|minimal|low|medium|high|xhigh|max with no toggle/budget—anti-pattern full enum and wrong baseline. Action: Match lab/peer shape for the completions path Zenifra serves; only add levels with host-specific proof.
• [high] [violation] providers/zenifra/models/qwen3.8-flash.toml:6 - Check: Same Qwen3.8 Flash baseline. Why: Lab is toggle + low/medium/xhigh + budget_tokens; this file dumps the full effort enum only. Action: Align to lab/peer controls for this host.
• [high] [violation] providers/zenifra/models/qwen3.8-27b.toml:6 - Check: Qwen3.8-27B peer/lab effort is low/medium/xhigh (often with toggle), not the full enum. Why: Cortecs/Cloudflare peers use low|medium|xhigh (± toggle). This file sets none|minimal|low|medium|high|xhigh|max. Action: Replace with the verified host set based on lab/peers, not the entire effort vocabulary.
• [medium] [possible mistake] providers/zenifra/models/kimi-k2.5.toml:7 - Check: [] on a relay requires affirmative “no caller control.” Why: Lab metadata is reasoning = true; many same-model relays use toggle. Empty here has no wire comment affirming no control—only a models-list citation. Action: Confirm whether Zenifra exposes thinking on/off (or effort); if it does, author that option; if not, state that explicitly in the leading comment.
• [medium] [possible mistake] providers/zenifra/models/qwen3.8-27b.toml:4 - Check: Overrides must stay consistent with inherited modalities. Why: Several entries set attachment = false (and/or drop text-only [modalities]) while base_model still supplies image/video (and pdf for max). Clients can inherit multimodal input with attachment = false. The prior Zenifra Qwen route overrode modalities to text-only. Action: For each text-only host route (qwen3.8-27b, qwen3.8-flash, qwen3.8-max, qwen3.7-plus, kimi-k2.5, kimi-k2.7-code, kimi-k3), either override [modalities].input to what Zenifra actually accepts, or keep attachment/modalities aligned with real vision support.
• [medium] [possible mistake] providers/zenifra/models/deepseek-v4-pro.toml:4 - Check: structured_output overrides need host evidence. Why: Lab DeepSeek/GLM/Qwen3.8-27B mark structured_output = true; several Zenifra files force false (and Qwen3.7 forces true where lab omits it) with only a models-list source. Action: Verify against Zenifra docs/API behavior and keep the override only where the host truly differs; remove redundant restates of the base value.
• [medium] [possible mistake] providers/zenifra/models/qwen3.7-plus.toml:18 - Check: Context tier threshold should match the provider’s pricing boundary. Why: Lab and other hosts use tier size 256_000 for Qwen3.7 Plus; this PR uses 256_001. That is a suspicious off-by-one if the band starts at 256k. Action: Confirm the Zenifra BRL tier cutover and set tier.size to that exact token threshold.
• [low] [possible mistake] .pr-review/pull-request.json - Check: Material cost/capability data should cite direct provider pricing/docs mapped to claims. Why: Body cites conversion rate and GET /v1/models, but not a pricing/docs page mapping BRL→USD fields, tier, limits, or reasoning wire paths per model. Action: Add first-party Zenifra pricing/docs (or models payload excerpts) stating what each supports: USD costs after BRL/5.20, tier, limits, and reasoning controls.

[github-actions]

Action items

• [high] [violation] providers/zenifra/models/qwen3.8-27b.toml:8 - Check: Reasoning options must follow the lab/peer baseline for this model, not a full effort enum dump. Why: Lab Alibaba and established peers use a narrow set (typically toggle + low/medium/xhigh, or none/low/medium/xhigh), not none|minimal|low|medium|high|xhigh|max. Authored full-schema dumps misrepresent caller controls and break the audit baseline. The same pattern is on qwen3.8-flash.toml and qwen3.8-max.toml. Action: Replace with the intersection Zenifra actually accepts that matches lab/peer controls (and add a leading wire comment). Drop levels that only appear because the gateway schema lists every enum value.
• [high] [violation] providers/zenifra/models/qwen3.7-max.toml:7 - Check: On multi-model relays, reasoning_options = [] means no caller control, not “not listed / unsure.” Why: First-party Alibaba uses toggle + budget_tokens for Qwen3.7 Max/Plus; OpenRouter peers expose at least a toggle. Comments only cite absence from supported_parameters, which is not enough when the lab and peers expose controls (and Zenifra already documents reasoning_effort on sibling Qwen3.8 routes). Same issue on qwen3.7-plus.toml. Action: Author the controls Zenifra forwards (toggle and/or budget/effort per this host’s wire path), or prove requests with lab-style thinking params are rejected and keep [] with that evidence.
• [high] [violation] providers/zenifra/models/kimi-k2.5.toml:7 - Check: Provider overrides must stay consistent after base_model merge. Why: attachment = false is set while [modalities] is not overridden, so the resolved model still inherits multimodal inputs (image/video on Kimi/Qwen bases) with attachments disabled—contradictory catalog data. Same pattern on kimi-k2.7-code.toml, kimi-k3.toml, qwen3.7-plus.toml, qwen3.8-27b.toml, qwen3.8-flash.toml, and qwen3.8-max.toml. Action: If Zenifra is text-only, override [modalities].input = ["text"] (and keep attachment = false). If multimodal input is real, drop the false attachment override.
• [medium] [possible mistake] providers/zenifra/models/kimi-k2.5.toml:10 - Check: Relay reasoning options for Kimi K2.5 vs lab/peer surface. Why: OpenRouter/Vercel peers expose a reasoning toggle for this model; Zenifra uses [] only because supported_parameters omits a control. That can under-report a real on/off path on an OpenAI-compatible host. Action: Confirm whether a toggle (or other thinking field) is accepted on Zenifra; if yes, author it with a leading wire comment; if no, keep [] and note the failed control path.
• [medium] [possible mistake] providers/zenifra/models/kimi-k3.toml:13 - Check: Provider limit.output should be a real max completion size, not the context window. Why: Override sets output = 1_048_576 while lab metadata is context = 1_048_576 / output = 131_072. Using the full context as output is a common models-list mis-map and can inflate completion limits. Action: Verify Zenifra’s max output tokens for this route; if it matches lab (~131K) or another documented cap, correct the override.
• [low] [possible mistake] providers/zenifra/models/qwen3.7-plus.toml:18 - Check: Context tier threshold should match the provider’s pricing band start. Why: Tier uses size = 256_001 while Alibaba and OpenRouter peers use 256_000 for the same Qwen3.7 Plus band. An off-by-one threshold mis-applies the higher price band. Action: Confirm Zenifra’s exact tier boundary and align size (almost certainly 256_000 unless docs say otherwise).
• [low] [possible mistake] providers/zenifra/models/deepseek-v4-flash-0731.toml:2 - Check: Non-USD costs must convert with a noted rate/date from a real FX source. Why: Comments call R$5.20 = US$1.00 a “project rate” with no exchange-rate source; other providers cite ECB/CBR (or similar). An unanchored rate can systematically skew all USD/MTok values. Action: Cite the FX source used for 2026-09-12 (or recompute from that source) in the leading conversion comment / PR body.

[github-actions]

Action items

• [high] [violation] providers/zenifra/models/zenifra/deepseek-v4-pro.toml:8 - Check: Reasoning options must match the lab / same-surface baseline for that model (not a generic enum). Why: First-party DeepSeek Pro and solid peers use toggle + effort high|max only (lab maps low/medium→high). This file and the new generate test both advertise low|high|max, inventing a Flash-style low level for Pro. Action: Change Pro effort to ["high", "max"], update the wire comment, and fix the Zenifra exposes only native reasoning controls expectation for zenifra/deepseek-v4-pro.
• [medium] [possible mistake] providers/zenifra/models/zenifra/glm-5.2.toml:9 - Check: Relay reasoning controls must match lab/peers for this host’s real wire surface. Why: ZhipuAI/Z.AI catalog GLM-5.2 as effort-only high|max (off via effort mapping, not a separate toggle). This PR adds { type = "toggle" } with enable_thinking, which is not the lab baseline. Action: Confirm Zenifra actually exposes enable_thinking for this model; if not, drop the toggle and keep effort high|max only (and update the test).
• [medium] [possible mistake] providers/zenifra/models/zenifra/deepseek-v4-flash-0731.toml:3 - Check: Toggle wire comments must name the real request field on this host. Why: Comments claim enable_thinking = true|false, while DeepSeek first-party and peers document thinking.type = enabled|disabled. A wrong wire path misleads consumers even if effort values are right. Action: Verify Zenifra’s DeepSeek toggle field and align the leading comments (and keep Flash effort as low|high|max only if that matches the host).
• [medium] [possible mistake] providers/zenifra/models/zenifra/qwen3.7-plus.toml:22 - Check: Context tier size is the threshold where the higher band starts. Why: Alibaba and established peers use 256_000 for Qwen3.7 Plus; this file uses 256_001, which shifts the tier boundary by one token versus the usual band. Action: Confirm Zenifra’s priced breakpoint; if it is the standard 256k band, set size = 256_000.
• [low] [possible mistake] .pr-review/pull-request.json (PR body) - Check: Data-changing PRs should cite direct provider pricing/docs mapped to claims. Why: Body notes BRL→USD at R$5.20 and the models list, but does not point to Zenifra pricing/docs pages or map which BRL list prices support each cost/tier. Action: Add first-party pricing/docs links and briefly state what each supports (list prices, tier breakpoint, model IDs / zenifra/ namespace).

[github-actions]

Action items

• [high] [violation] providers/zenifra/models/zenifra/qwen3.8-27b.toml:6 - Check: Provider overrides after base_model must stay consistent with inherited modalities (real host deltas only). Why: The file sets attachment = false but does not override [modalities]. The lab base (models/alibaba/qwen3.8-27b.toml) still contributes input = ["text", "image", "video"] and attachment = true, so the merged model advertises vision/video inputs while claiming no attachments. Action: Either drop attachment = false (and keep multimodal inheritance) or keep text-only hosting and set [modalities] input = ["text"] (and matching output) alongside attachment = false. Update the new attachment test expectation to match.
• [medium] [possible mistake] providers/zenifra/models/zenifra/kimi-k3.toml:9 - Check: Provider limit.output must reflect the host’s real max output, not context. Why: output = 1_048_576 matches the lab context size, while lab/Moonshot metadata use output = 131_072. Treating full context as max output is a common catalog error and inflates capability. Action: Verify Zenifra’s advertised max completion tokens for zenifra/kimi-k3; if it is not ~1M, set limit.output to the real value (likely ~131_072) or omit the override to inherit the lab default.
• [medium] [possible mistake] providers/zenifra/models/zenifra/qwen3.7-plus.toml:18 - Check: Context pricing tiers should use the provider’s real threshold. Why: tier.size = 256_001 is unique in-tree; Alibaba and established relays use 256_000 for Qwen3.7 Plus long-context pricing. An off-by-one threshold mis-bands cost. Action: Confirm the exact Zenifra/BRL tier cutoff from pricing docs; if the band starts at 256K tokens, use size = 256_000 (or document why 256_001 is intentional).
• [low] [possible mistake] providers/zenifra/models/zenifra/deepseek-v4-pro.toml:7 - Check: Relay reasoning effort must match lab/same-surface peers unless this host proves extra levels. Why: First-party DeepSeek V4 Pro is toggle + high|max only; this entry adds low (Flash-style). Extra levels need host-proven effect, not only a /v1/models assertion. Action: Confirm Zenifra actually accepts and applies reasoning_effort=low for zenifra/deepseek-v4-pro; if not, use ["high", "max"] like providers/deepseek/models/deepseek-v4-pro.toml.

[github-actions]

Action items

• [high] [possible mistake] providers/zenifra/models/zenifra/kimi-k3.toml:8 - Check: Provider [limit] overrides must reflect the host's real serving limits, not swap context/output. Why: File sets [limit] output = 1_048_576 while models/moonshotai/kimi-k3.toml has context = 1_048_576 / output = 131_072; provider output equals lab context and is 8x lab output, indicating context was copied into output. Action: Correct to a verified host output limit or remove the override to inherit 131_072.
• [high] [violation] providers/zenifra/models/zenifra/glm-5.1.toml:3 - Check: Every toggle requires a leading top-of-file wire comment with the exact request field (AGENTS.md → Reasoning options). Why: glm-5.1.toml, qwen3.6-flash.toml, and qwen3.6-plus.toml author reasoning_options = [{ type = "toggle" }] but only have # No native reasoning_effort levels advertised; expose the model's thinking toggle only with no Toggle: ... wire path, unlike the compliant # Toggle: enable_thinking = true|false used elsewhere in this PR. Action: Add the exact Zenifra wire path for the toggle to the header block in all three files.
• [medium] [possible mistake] providers/zenifra/models/zenifra/kimi-k2.5.toml:4 - Check: base_model files must keep only real provider deltas, not restate or invent capabilities. Why: kimi-k2.5.toml, kimi-k2.7-code.toml, and kimi-k3.toml set temperature = true while models/moonshotai/kimi-k2.5.toml, kimi-k2.7-code.toml, and kimi-k3.toml all have temperature = false and providers/moonshotai/models/kimi-k3.toml / kimi-k2.7-code.toml inherit false; a relay cannot make a model respect temperature the lab marks unsupported. Action: Remove the temperature = true overrides or provide direct Zenifra evidence that temperature has effect on these routes.
• [medium] [possible mistake] providers/zenifra/models/zenifra/deepseek-v4-pro.toml:5 - Check: Relay effort sets must copy the lab + same-surface peer baseline, not narrow without host evidence. Why: First-party providers/deepseek/models/deepseek-v4-pro.toml is toggle + effort [low, high, max] and Zenifra deepseek-v4-flash-0731.toml in this PR keeps [low, high, max], but Zenifra deepseek-v4-pro.toml narrows to [high, max] with only a GET /v1/models assertion; same host differing per-model needs verification. Action: Verify Zenifra supported_parameters for this route and restore low if the host forwards it.
• [medium] [possible mistake] providers/zenifra/models/zenifra/glm-5.2.toml:5 - Check: Do not add or drop toggle controls relative to the native baseline without host evidence. Why: First-party providers/zhipuai/models/glm-5.2.toml is effort [high, max] only with off via none|minimal, but Zenifra adds { type = "toggle" }; conversely first-party providers/moonshotai/models/kimi-k3.toml is toggle + effort [low, high, max] but Zenifra kimi-k3.toml drops the toggle as always-on. Action: Confirm from Zenifra supported_parameters/docs whether enable_thinking exists for GLM-5.2 and whether Kimi-K3 has no off control, and align to the verified surface.

[github-actions]

Action items

• [critical] [violation] providers/zenifra/models/zenifra/deepseek-v4-pro.toml:1 - Check: Relay model IDs must preserve the lab/upstream namespace, not use a provider namespace. Why: Base revision uses providers/zenifra/models/alibaba/qwen3.6-35b-a3b.toml (alibaba/...) and AGENTS.md/OpenRouter/Kilo use providers/<relay>/models/<lab>/... (e.g. providers/openrouter/models/anthropic/...); this PR moves all routes to providers/zenifra/models/zenifra/... yielding zenifra/... IDs that imply a zenifra lab and break AI-SDK lookup consistency. Action: Restore lab-namespaced paths (e.g. providers/zenifra/models/deepseek/..., providers/zenifra/models/alibaba/..., providers/zenifra/models/moonshotai/..., providers/zenifra/models/zhipuai/...) matching the base_model lab, or provide API evidence that Zenifra serves zenifra/... IDs.
• [high] [violation] packages/core/test/generate.test.ts:805 - Check: Catalog PRs are catalog-only (AGENTS.md). Why: This PR adds provider-specific snapshot tests (Zenifra model IDs..., Zenifra exposes only native reasoning controls, Zenifra attachment..., Zenifra serves current limits...) that lock in its own contested IDs/costs/limits/reasoning values and will fail legitimate future corrections. Action: Revert all packages/core/test/generate.test.ts changes; keep the PR to models/ + providers/ data files.
• [high] [possible mistake] providers/zenifra/models/zenifra/kimi-k3.toml:10 - Check: [limit] correctness vs inherited lab limits. Why: Diff authors [limit] output = 1_048_576 while models/moonshotai/kimi-k3.toml is context = 1_048_576, output = 131_072; merged result becomes context 1M/output 1M (8x lab output, output equals context), with no comment or tier justification. Action: Verify Zenifra's advertised output limit; if not 1M, remove the [limit] override to inherit lab limits or correct to the documented context/output values.
• [high] [possible mistake] providers/zenifra/models/zenifra/glm-5.2.toml:6 - Check: Reasoning baseline must copy lab + same-surface peers; do not invent toggle. Why: First-party providers/zhipuai/models/glm-5.2.toml is effort-only ["high","max"] with no toggle, but Zenifra authors toggle + effort ["high","max"] with only enable_thinking assertion. Action: Remove the invented toggle or cite Zenifra docs/supported_parameters showing a real on/off field forwarded on this route.
• [medium] [possible mistake] providers/zenifra/models/zenifra/deepseek-v4-pro.toml:8 - Check: Baseline effort must be native/peer set, consistent per family on same host. Why: First-party providers/deepseek/models/deepseek-v4-pro.toml and OpenRouter peers use toggle + ["low","high","max"]; Zenifra flash keeps low|high|max but pro drops to high|max on the same host with only a GET /v1/models comment, an unexplained per-model divergence. Action: Align both to the lab set or provide per-model Zenifra evidence that low is unsupported on pro but supported on flash.
• [medium] [possible mistake] providers/zenifra/models/zenifra/kimi-k3.toml:5 - Check: Do not drop lab toggle from uncertainty. Why: First-party providers/moonshotai/models/kimi-k3.toml is toggle + effort ["low","high","max"]; Zenifra keeps effort but drops toggle claiming “always on” without showing supported_parameters/docs that off is unavailable. Action: Restore toggle with wire comment if Zenifra forwards thinking.type, or cite affirmative host evidence that thinking cannot be disabled.
• [medium] [possible mistake] providers/zenifra/models/zenifra/qwen3.8-max.toml:6 - Check: budget_tokens only for real reasoning budgets; do not silently drop lab budget. Why: First-party providers/alibaba/models/qwen3.8-max.toml / qwen3.8-flash.toml / qwen3.7-max / qwen3.7-plus / qwen3.6-flash / qwen3.6-plus all expose budget_tokens (thinking_budget); all Zenifra Qwen routes drop it to toggle or toggle+effort with “no native effort levels” notes but no statement that thinking_budget is absent on Zenifra. Action: Verify whether Zenifra forwards thinking_budget; restore {type="budget_tokens"} with bounds if it does, or document the absence in the leading header.
• [low] [violation] .pr-review/diff.patch:4 - Check: Data-changing PRs should cite direct provider pricing/docs/API mapped to claims. Why: PR body lists conversions and capability changes but provides no URLs; TOML headers cite only GET https://ai.zenifra.com/v1/models which cannot be verified here, leaving expanded limits (deepseek output 393_216 vs lab 384_000, qwen3.7-max/plus output 131_072 vs lab 65_536/64_000, glm-5.1 context 202_745 vs 200_000, glm-5.2 context 1_048_576 vs 1_000_000, tiers 32_001/256_001) and narrowed temperature/structured_output/modalities unreviewable. Action: Add first-party or Zenifra doc/API references stating what each citation supports, and clarify which price band (peak vs noted off-peak BRL rows) the USD costs represent.

[ramonpaolo]

@rekram1-node

Hi

Can you merge?